mcp-sage

An MCP (Model Context Protocol) server that provides tools for sending prompts to OpenAI's GPT-5, GPT-4.1, Google's Gemini 2.5 Pro, or Anthropic's Claude Opus 4.1 based on token count and configuration. The tools embed all referenced filepaths (recursively for folders) in the prompt. This is useful for getting second opinions or detailed code reviews from models that can handle large amounts of context accurately.

Rationale

I make heavy use of Claude Code. It's a great product that works well for my workflow. Newer models with large amounts of context seem really useful though for dealing with more complex codebases where more context is needed. This lets me continue to use Claude Code as a development tool while leveraging the large context capabilities of GPT-5, Gemini 2.5 Pro, and other models to augment Claude Code's limited context.

Model Selection

The server automatically selects the appropriate model based on token count, with configuration defined in models.yaml:

- For smaller contexts (≤ 400K tokens): Uses OpenAI's GPT-5 (if OPENAI_API_KEY is set)
- For medium contexts (≤ 1M tokens): Uses Google's Gemini 2.5 Pro (if GEMINI_API_KEY is set)
- For fallback (≤ 1M tokens): Uses OpenAI's GPT-4.1
- If the content exceeds 1M tokens: Returns an informative error

Fallback behavior:
- API Key Fallback:
- If OPENAI_API_KEY is missing, Gemini will be used for all contexts within its 1M token limit
- If GEMINI_API_KEY is missing, only smaller contexts can be processed with OpenAI models
- If required API keys are missing, an informative error is returned

Inspiration

This project draws inspiration from two other open source projects:

- simonw/files-to-prompt for the file compression
- asadm/vibemode for the idea and prompt to send the entire repo to Gemini for wholesale edit suggestions
- PhialsBasement/Chain-of-Recursive-Thoughts inspiration for the debate functionality

Overview

This project implements an MCP server that exposes two primary tools:

sage-opinion

1. Takes a prompt and a list of file/dir paths as input
2. Packs the files into a structured XML format
3. Measures the token count and selects the appropriate model:
- GPT-5 for ≤ 400K tokens
- Gemini 2.5 Pro for > 400K and ≤ 1M tokens
- GPT-4.1 as fallback for ≤ 1M tokens
4. Sends the combined prompt + context to the selected model
5. Returns the model's response

sage-review

1. Takes an instruction for code changes and a list of file/dir paths as input
2. Packs the files into a structured XML format
3. Measures the token count and selects the appropriate model:
- GPT-5 for ≤ 400K tokens
- Gemini 2.5 Pro for > 400K and ≤ 1M tokens
- GPT-4.1 as fallback for ≤ 1M tokens
4. Creates a specialized prompt instructing the model to format responses using SEARCH/REPLACE blocks
5. Sends the combined context + instruction to the selected model
6. Returns edit suggestions formatted as SEARCH/REPLACE blocks for easy implementation

Debate Mode

Both sage-opinion and sage-review support an optional debate mode that can be enabled by adding debate: true to the arguments. When enabled, the system orchestrates a structured debate between multiple models to generate higher-quality responses.

--------------------------------------------------------------------

1. Multi-Model Debate Flow

flowchart TD

  S0[Start Debate] -->|determine models, judge, budgets| R1

subgraph R1["Round 1"]

    direction TB

    R1GEN["Generation Phase<br/>ALL models run in parallel"]

    R1GEN --> R1CRIT["Critique Phase<br/>ALL models critique others in parallel"]

  end
subgraph RN["Rounds 2 to N"]

    direction TB

    SYNTH["Synthesis Phase<br/>every model refines own plan"]

    SYNTH --> CONS[Consensus Check]

    CONS -->|Consensus reached| JUDGE

    CONS -->|No consensus & round < N| CRIT["Critique Phase<br/>models critique in parallel"]

    CRIT --> SYNTH

  end
R1 --> RN

  JUDGE[Judgment Phase<br/>judge model selects/merges response]

  JUDGE --> FP[Final Response]
classDef round fill:#e2eafe,stroke:#4169E1;

  class R1GEN,R1CRIT,SYNTH,CRIT round;

  style FP fill:#D0F0D7,stroke:#2F855A,stroke-width:2px

  style JUDGE fill:#E8E8FF,stroke:#555,stroke-width:1px

Key phases in the multi-model debate:

Setup Phase
- The system determines available models, selects a judge, and allocates token budgets

Round 1
- Generation Phase - Every available model (A, B, C, etc.) generates its response in parallel
- Critique Phase - Each model reviews all other responses (never its own) and produces structured critiques in parallel

Rounds 2 to N (N defaults to 3)
1. Synthesis Phase - Each model improves its previous response using critiques it received (models work in parallel)
2. Consensus Check - The judge model scores similarity between all current responses
- If score ≥ 0.9, the debate stops early and jumps to Judgment
3. Critique Phase - If consensus is not reached AND we're not in the final round, each model critiques all other responses again (in parallel)

Judgment Phase
- After completing all rounds (or reaching early consensus), the judge model (Claude Opus 4.1 by default):
- For sage-opinion: Selects the single best response (no synthesis)
- For sage-review: Can either select the best response OR merge multiple responses
- Provides a confidence score for its selection/synthesis

--------------------------------------------------------------------

2. Self-Debate Flow - Single Model Available

flowchart TD

  SD0[Start Self-Debate] --> R1

subgraph R1["Round 1 - Initial Responses"]

    direction TB

    P1[Generate Response 1] --> P2[Generate Response 2<br/>different approach]

    P2 --> P3[Generate Response 3<br/>different approach]

  end
subgraph RN["Rounds 2 to N"]

    direction TB

    REF[Generate Improved Response<br/>addresses weaknesses in all previous responses]

    DEC{More rounds left?}

    REF --> DEC

    DEC -->|Yes| REF

  end
R1 --> RN

  DEC -->|No| FP[Final Response = last response generated]
style FP fill:#D0F0D7,stroke:#2F855A,stroke-width:2px

When only one model is available, a Chain of Recursive Thoughts (CoRT) approach is used:

1. Initial Burst - The model generates three distinct responses, each taking a different approach
2. Refinement Rounds - For each subsequent round (2 to N, default N=3):
- The model reviews all previous responses
- It critiques them internally, identifying strengths and weaknesses
- It produces one new improved response that addresses limitations in earlier responses
3. Final Selection - The last response generated becomes the final output

--------------------------------------------------------------------

What Actually Happens in Code (quick reference)

| Phase / Functionality | Code Location | Notes |
|-----------------------|---------------|-------|
| Generation Prompts | prompts/debatePrompts.generatePrompt | Creates initial responses from each model |
| Critique Prompts | prompts/debatePrompts.critiquePrompt | Uses "## Critique of {ID}" sections |
| Synthesis Prompts | prompts/debatePrompts.synthesizePrompt | Model revises its own response |
| Consensus Check | orchestrator/debateOrchestrator | Judge model returns JSON with consensusScore |
| Judgment | prompts/debatePrompts.judgePrompt | Judge returns final response + confidence |
| Self-Debate Prompt | prompts/debatePrompts.selfDebatePrompt | Chain-of-Recursive-Thoughts loop |

Performance and Cost Considerations

⚠️ Important: When using debate mode:
- It can take more time to complete (2-5 minutes with multiple models)
- Consumes more API tokens due to multiple rounds of debate
- Incurs higher costs than single-model approaches

Typical resource usage:
- Multi-model debate: 2-4x more tokens than a single model approach
- Processing time: 2-5 minutes depending on complexity and model availability
- API costs vary by models used and complexity

Prerequisites

- Node.js (v18 or later)
- API keys for the models you want to use:
- OpenAI API key (for GPT-5 and GPT-4.1)
- Google Gemini API key (for Gemini 2.5 Pro)
- Anthropic API key (for Claude Opus 4.1 as judge in debates)

Note: While the server can function with just one API key, it works best when all three are provided. This enables:
- Optimal model selection based on token count
- Multi-model debates for higher quality responses
- Claude Opus 4.1 as an impartial judge in debate mode

Installation

Installing via Smithery

To install Sage for Claude Desktop automatically via Smithery:

npx -y @smithery/cli install @jalehman/mcp-sage --client claude

Installing manually

```bash